Report Grader: AI grading by Fig AI Systems

It sits alongside Build-a-Dress, a live business we run at a public address. Both are set out on the products page.

Consistency is the other problem. The same report can come back with different emphasis and different feedback from two reviewers. The same reviewer can differ on two different days. The standards are long, and a human reader holds only part of them in mind. Both problems get worse as volume rises.

Report Grader takes on the repetitive half of that job: the systematic comparison against the standards.

Three decisions define the architecture:

• The unit of work is a segment, not the whole file. That makes the analysis parallelisable and each judgment attributable.
• The grading agents are mapped to a domain-specific corpus, so criteria come from the actual review standards.
• Evaluation is a permanent part of the product - the same loop described under closed-loop AI evaluation.

The fan-out costs more calls and more orchestration. It buys precision, concurrency and the ability to inspect any single grading decision. Managing it is what multi-agent orchestration and agentic workflow automation exists to do.

DOCX and PDF parsing with segmentation

Reports arrive as DOCX and PDF files. The first stage parses them into text, then segments it into pieces small enough to analyse independently. Segmentation preserves the structure of the report, so a grading decision can be pointed back at its source.

1,000+ token-efficient LLM calls at high concurrency

A single report analysis triggers more than 1,000 LLM calls. That works because the calls are token-efficient and run in parallel. Each call carries the smallest context that will do its job, which keeps cost per report in proportion. Concurrency keeps the wall-clock time of a thousand-call analysis acceptable to someone waiting for a result.

Agents mapped to a domain-specific corpus

The grading agents are skill-enabled and mapped directly to a domain-specific corpus. Grading is grounded in the school review criteria themselves. This is a retrieval problem: getting the right slice of authoritative material in front of the model. That is the job of vector databases, retrieval and data clustering.

Evaluation patterns that close the loop

Evaluation patterns close the loop between model output, review criteria and system improvement. Without that loop, outputs get spot-checked and changes get made on instinct. With it closed, a disagreement between the system and the criteria becomes a signal for the next iteration.

Human-in-the-loop grading

Inspectors can put their own input into the grading where it is needed. The product does the exhaustive first pass. A person applies judgment where the situation calls for it. The output affects a school, so that gate was built in from the start.

The one number on this page, the thousand-plus calls per analysis, describes how the system is built. Any figure for time saved depends on the standard being graded and the team doing the grading, so it belongs in a conversation about your reports and not on a marketing page.

The evaluation work is closed-loop AI evaluation and testing. The corpus grounding is vector databases, retrieval and clustering. Reassembling a thousand scoped calls into one graded document is multi-agent orchestration. For a consumer marketplace, see Build-a-Dress or all products.